An emergency power repair engineering scheme decision method and system

Abstract

The application discloses an emergency power repair engineering scheme decision method and system, and relates to the field of power repair. The existing power emergency repair lacks full-process dynamic cost accounting, multi-dimensional collaborative decision and small sample intelligent decision-making capability. The application comprises the following steps: constructing a full-process cost accounting model, including a direct engineering cost and an emergency measure cost accounting sub-model, introducing a dynamic correction factor to adapt to the uncertainty of the repair scene; establishing a decision index system including an economic benefit, a repair progress and a field safety criterion layer, and clearly defining the positive and negative attributes of each index; constructing a decision model based on an improved convolutional neural network, using an SGDM optimizer and an L2 regularization term, and constructing training samples by dividing the index value interval; inputting the index vector of the alternative scheme into the trained model to obtain a decision value, and selecting the optimal repair scheme according to the decision value. The technical scheme realizes full-process dynamic cost accounting and multi-objective collaborative decision-making, and improves the scientificity and accuracy of scheme optimization.

Description

Technical Field

[0001] This invention relates to the field of power emergency repair, and in particular to a method and system for decision-making on power emergency repair engineering schemes. Background Technology

[0002] Against the backdrop of frequent extreme weather events and large-scale grid integration of renewable energy, the risk of damage to power grid equipment has intensified. Rapid response and scientific decision-making in power emergency repair projects have become crucial to ensuring power supply reliability. Power emergency repair is characterized by urgency, coordination, specialization, and high risk. Its project cost includes not only conventional direct engineering costs but also substantial emergency measures expenses, and is significantly affected by dynamic factors such as terrain, climate, and equipment condition. However, existing cost accounting for power emergency repair projects mostly uses static quota methods, lacking dynamic accounting models covering the entire process from pre-event to during-event to post-event. This leads to a serious disconnect between budget preparation and actual settlement, making it difficult to accurately reflect the true consumption of repair resources.

[0003] In terms of decision-making, existing research mostly focuses on single-dimensional optimization, such as considering only repair costs or technical feasibility, and has not yet constructed a multi-dimensional collaborative decision-making indicator system that integrates economy, timeliness, and safety. At the same time, traditional multi-attribute decision-making methods (such as the analytic hierarchy process and fuzzy comprehensive evaluation) rely on manual weighting, which is highly subjective and difficult to handle complex nonlinear relationships between indicators; while conventional neural network models are prone to overfitting in small sample and multi-indicator scenarios, and have insufficient generalization ability, which limits their application in emergency repair scenarios with high randomness and scarce samples.

[0004] Therefore, there is an urgent need for an intelligent decision-making system that can achieve dynamic cost accounting throughout the entire process, multi-dimensional collaborative evaluation, and small-sample learning capabilities, so as to provide support for the scientific selection of power emergency repair engineering schemes. Summary of the Invention

[0005] The technical problem to be solved and the technical task proposed by this invention is to improve and refine existing technical solutions, and to provide a decision-making method and system for power emergency repair engineering schemes, so as to achieve dynamic cost accounting throughout the entire process, multi-dimensional collaborative evaluation, and intelligent decision-making under small sample conditions. To this end, this invention adopts the following technical solution.

[0006] The first aspect: providing a decision-making method for power emergency repair engineering schemes, including the following steps: 1) Construct a full-process cost accounting model for power emergency repair projects. The full-process cost accounting model includes a direct engineering cost accounting sub-model and an emergency measures cost accounting sub-model, which are used to calculate the cost of alternative repair schemes to obtain direct engineering costs and emergency measures costs. Among them, at least one dynamic correction factor is introduced in the cost accounting to adapt to the uncertainty of the repair scenario. 2) Establish a decision-making indicator system for power emergency repair engineering schemes. The indicator system includes three criterion layers: economic benefits, repair progress, and on-site safety. Each criterion layer has multiple operational level indicators, and the positive or negative attributes of each operational level indicator are clearly defined. 3) Obtain multiple alternative solutions to be decided, calculate the cost data of each alternative solution according to the cost accounting model, and determine the value of each alternative solution on each operational level indicator according to the indicator system to form a solution indicator vector. 4) Construct a scheme decision model based on an improved convolutional neural network, use the SGDM optimizer to optimize the network parameters, and introduce an L2 regularization term into the loss function; 5) Construct a training sample set, which is obtained by equally dividing the value range of each index. Each training sample corresponds to a decision value. Use the training sample set to train the improved convolutional neural network. 6) Input the index vectors of the alternative schemes into the trained improved convolutional neural network to obtain the decision values ​​of each scheme, and select the optimal emergency repair scheme based on the magnitude of the decision values.

[0007] This technical solution constructs a full-process cost accounting model that includes direct engineering costs and emergency measures costs, and introduces a dynamic correction factor into the accounting process. This allows for dynamic adjustment of costs based on uncertainties such as terrain, machine conditions, and climate, making the accounting results closer to actual resource consumption. This solves the problem of the disconnect between traditional static quotas and actual settlement, providing an accurate cost data foundation for subsequent scheme decisions. By establishing a decision indicator system encompassing three criterion layers—economic benefits, repair progress, and on-site safety—and clarifying the positive or negative attributes of each indicator, a multi-objective collaborative evaluation of cost, timeliness, and safety is achieved. This avoids decision-making biases that sacrifice repair efficiency or safety by simply pursuing low cost, and better meets the actual needs of power emergency repair. By introducing a convolutional neural network to construct the decision model, the nonlinear mapping relationship between decision indicators and the comprehensive evaluation value of the scheme is automatically learned, avoiding the subjectivity of human weighting. At the same time, convolutional kernels are designed for the characteristics of one-dimensional indicator sequences, which can effectively mine the local correlation features of adjacent indicators under the same criterion layer, making the decision results more scientific. The SGDM optimizer, by introducing a momentum term, effectively mitigates gradient oscillations and accelerates convergence. Simultaneously, an L2 regularization term is introduced into the loss function to suppress overfitting by penalizing excessively large weights. The synergistic effect of these two methods allows the model to maintain good generalization performance even with limited training samples. By constructing training samples by equally dividing the value ranges of each indicator, the optimal and worst values ​​of the indicators in the alternative schemes are used to determine the value range. Multiple virtual samples are generated after equal division and assigned corresponding decision values, effectively expanding the training sample set and providing a data foundation for training the deep learning model. This technical solution organically integrates full-process cost accounting, multi-dimensional indicator system construction, intelligent decision model training, and scheme optimization, forming a complete technical chain from cost data input to optimal scheme output. Decision-makers can complete the entire process from scheme cost calculation to comprehensive evaluation without switching between multiple tools or methods, improving the efficiency and scientific rigor of emergency repair scheme decision-making.

[0008] As a preferred technical means: the direct engineering cost accounting sub-model includes construction machinery usage cost accounting, material cost accounting, and labor cost accounting; The construction machinery usage fee is determined based on the amount of machinery used and the unit price of machinery per shift, wherein the amount of machinery used is calculated based on the amount of work, the machinery shift operation rate, the work shift system and the planned construction period. The material cost includes priced material cost and unpriced material cost. The fuel cost in the unpriced material cost is corrected by a dynamic fuel consumption correction factor, which includes a speed coefficient, an engine condition coefficient, and a condition coefficient. The labor cost is determined based on the number of personnel and the unit price of labor.

[0009] This technical solution refines the calculation components of direct engineering costs, clarifying the specific calculation methods for construction machinery usage fees, material costs, and labor costs. Specifically, it introduces dynamic fuel consumption correction factors (speed coefficient, machine condition coefficient, and condition coefficient) for fuel costs. This allows for dynamic adjustments to fuel consumption based on actual working conditions at the repair site, such as machinery speed, technical condition, and altitude. This makes material cost calculations more aligned with actual site conditions, improving the accuracy and adaptability of cost accounting.

[0010] As a preferred technical means, the fuel consumption dynamic correction factor is used to correct the basic fuel consumption quota of the machinery in order to calculate the actual fuel consumption of the machinery under emergency repair conditions.

[0011] This technical solution clarifies the specific target of the dynamic fuel consumption correction factor, namely, to correct the basic fuel consumption quota for machinery, thereby obtaining the actual fuel consumption. This correction process transforms the static quota into a dynamic value, effectively solving the problem of the disconnect between traditional quotas and emergency repair conditions, and providing a technical means for accurate calculation of fuel costs.

[0012] As a preferred technical means: the emergency measures cost accounting sub-model is used to calculate emergency dispatch and coordination costs, safe and civilized construction and protection measures costs, and construction operation increase costs. Each cost is calculated according to the consumption and unit price of labor costs, material costs, and machinery usage costs.

[0013] This technical solution establishes a classification and accounting framework for emergency response costs, incorporating unique expenses such as emergency dispatch and coordination, safety protection, and increased construction work into a unified model. These costs are calculated separately according to three basic elements: labor, materials, and machinery. This makes the accounting of emergency response costs more systematic and standardized, providing complete cost data support for subsequent decision-making.

[0014] As a preferred technical means: the operational level indicators mentioned in step 2) include: power emergency engineering measures cost, direct engineering cost, value change ratio of fixed assets before and after emergency repair, return on investment, time required for power supply to return to normal, time from receiving the fault alarm to the arrival of personnel, materials and equipment on site, time from arrival on site to the start of emergency repair, the difference between actual emergency repair time and planned emergency repair time, the proportion of emergency repairs using the latest technology and equipment, the proportion of emergency repair efficiency improvement due to the introduction of new equipment or technology, the frequency of unexpected events during emergency repair, the proportion of emergency repair personnel who have received safety training, the degree of implementation of the emergency plan in actual emergency repair, and the evaluation score of the effect of emergency plan implementation. Among them, the following are negative indicators: power emergency engineering measures cost, direct engineering cost, time required to restore power supply to normal status, time from receiving a fault alarm to the arrival of personnel, materials and equipment on site, time from arrival on site to the start of emergency repair, the ratio of actual repair time to planned repair time, the proportion of repairs using the latest technology and equipment, and the frequency of unexpected events during the repair process; the following are positive indicators: the ratio of fixed asset value change before and after repair, return on investment, the proportion of improvement in repair efficiency due to the introduction of new equipment or technology, the proportion of repair personnel receiving safety training, the degree of implementation of the emergency plan in actual repairs, and the evaluation of the effectiveness of the emergency plan after implementation.

[0015] This technical solution constructs a detailed decision-making indicator system comprising 14 operational-level indicators, comprehensively covering three dimensions: economic benefits (cost, revenue, recovery speed), emergency repair progress (response time, repair time, technology updates), and on-site safety (accident frequency, training coverage, contingency plan execution). By clarifying the positive / negative attributes of each indicator, it provides standardized input for the subsequent training of the convolutional neural network model, ensuring that the model can accurately identify the direction of indicator advantages and disadvantages, and achieve scientific decision-making through multi-objective collaborative optimization.

[0016] As a preferred technical means: the method for constructing training samples in step 5) is as follows: for each indicator, the value range of the indicator is determined according to the optimal and worst values ​​in the alternative schemes, the value range is divided into several parts to obtain multiple interval points, each interval point corresponds to an indicator value, the interval points of all indicators are combined to obtain multiple training samples, and the corresponding decision values ​​are assigned in order from optimal to worst, with the decision value range being [0,1].

[0017] This technical solution constructs training samples by equally dividing the index intervals. It uses the optimal and worst values ​​from the alternative solutions to determine the value range of each index. By generating virtual samples through equal division and assigning corresponding decision values, it effectively expands the training sample set and solves the problem of limited historical data in the emergency repair field, which makes it difficult to support deep learning training. Simultaneously, the samples cover the complete gradient from optimal to worst, enabling the model to learn the continuous mapping relationship between index changes and decision values.

[0018] As a preferred technical means: the improved convolutional neural network includes an input layer, a convolutional layer, an activation layer, a fully connected layer, and an output layer; the number of nodes in the input layer of the improved convolutional neural network is equal to the number of indicators in the operation layer, the number of nodes in the output layer is 1, the convolutional layer uses multiple convolutional kernels of fixed size, the activation function is the ReLU function, and there are two fully connected layers.

[0019] This technical solution designs a convolutional neural network structure suitable for one-dimensional decision index sequences. The number of nodes in the input layer corresponds to 14 indicators, and the output layer is a single decision value. The convolutional layers use fixed-size convolutional kernels, which can automatically learn the local correlation features between adjacent indicators (such as indicators under the same criterion layer) during the sliding process. The ReLU activation function introduces nonlinearity. Two fully connected layers further integrate features. This fully adapts to the inherent characteristics of multi-indicator decision problems and effectively improves the expressive power of the model.

[0020] As a preferred technical means: the loss function of the improved convolutional neural network is a loss function with added L2 regularization term. The loss function with added L2 regularization term is the average loss of all training samples plus the L2 regularization penalty term multiplied by the regularization coefficient. When updating the network parameters using the SGDM optimizer, a momentum term is introduced during the update process of weights and biases to reduce gradient oscillations, and overfitting is suppressed by the L2 regularization term.

[0021] This technical solution addresses gradient oscillations and overfitting issues under limited sample training by combining the SGDM optimizer and L2 regularization. The SGDM optimizer introduces a momentum term, which utilizes historical gradient information to smooth parameter update directions, accelerating convergence and avoiding getting trapped in local optima. The L2 regularization term, by penalizing excessively large weights, limits model complexity and effectively suppresses overfitting. The combination of these two methods significantly improves the model's generalization ability and training stability with limited samples.

[0022] As a preferred technical means, the dynamic correction factor includes at least one of the terrain complexity adjustment coefficient, machine condition coefficient and condition coefficient, which is used to correct the construction machinery usage cost or material cost.

[0023] This technical solution clarifies the specific types of dynamic correction factors and their application objects. The terrain complexity adjustment coefficient reflects the impact of complex terrain such as mountains and swamps on construction efficiency; the machine condition coefficient reflects the impact of the age of mechanical equipment on fuel consumption and efficiency; and the condition coefficient reflects the impact of natural conditions such as altitude and climate on operations. By introducing these correction factors, cost accounting can dynamically respond to the diverse and complex working conditions at the emergency repair site, significantly improving the adaptability and accuracy of cost accounting.

[0024] The second aspect: Provides a decision-making system for power emergency repair engineering schemes, including: The cost accounting module is used to construct a full-process cost accounting model for power emergency repair projects, including direct engineering costs and emergency measures cost accounting sub-models. It performs cost accounting for alternative repair schemes and introduces at least one dynamic correction factor to adapt to the uncertainty of the repair scenario. The indicator system construction module is used to establish a decision-making indicator system for power emergency repair engineering schemes, which includes three criterion layers: economic benefits, repair progress, and on-site safety. Each criterion layer has several operational level indicators, and the positive or negative attributes of each operational level indicator are clearly defined. The scheme indicator acquisition module is used to acquire multiple alternative emergency repair schemes to be decided, call the cost accounting module to obtain the cost data of each scheme, and then determine the values ​​of each scheme's indicators at each operational level based on the decision indicator system to form a scheme indicator vector. The neural network building module is used to construct an improved convolutional neural network scheme decision model containing an input layer, convolutional layer, activation layer, fully connected layer and output layer. It uses the SGDM optimizer to optimize network parameters and introduces an L2 regularization term into the loss function. The model training module is used to construct a training sample set obtained by equally dividing the index value range, each training sample corresponding to a decision value, and to train the improved convolutional neural network using the training sample set. The scheme decision module is used to input the index vectors of each alternative scheme into the trained scheme decision model, obtain the decision value of each scheme, and select the optimal emergency repair scheme based on the magnitude of the decision value.

[0025] The cost accounting module of this technical solution includes sub-models for both direct engineering costs and emergency measures costs, and introduces a dynamic correction factor. This allows for dynamic adjustment of costs based on actual working conditions such as terrain complexity, machine conditions, altitude, and climate, overcoming the shortcomings of traditional static quota accounting which is disconnected from actual site conditions. This provides an accurate and reliable cost data foundation for subsequent decision-making, enabling the system to adapt to complex emergency repair scenarios such as mountainous areas, freezing rain, and high altitudes. The indicator system construction module establishes a decision indicator system encompassing three criterion layers: economic benefits, repair progress, and on-site safety. It clearly defines the positive and negative attributes of each indicator, thus comprehensively reflecting the overall benefits of emergency repair plans. This avoids decision-making biases that sacrifice repair efficiency or safety in the pursuit of low cost, making the system's decision results more aligned with the actual needs of power emergency management. The neural network construction module employs an improved convolutional neural network as the decision-making model. Through convolutional layers, it automatically learns the local correlation features between adjacent indicators, enabling the discovery of complex nonlinear relationships within and between dimensions of economic benefits, repair progress, and on-site safety. Compared to traditional methods such as the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation, which rely on manual weighting, this system achieves objective and adaptive intelligent decision-making, avoiding the interference of subjective factors on the decision results. Addressing the characteristics of emergency repair projects—highly unpredictable and with scarce historical samples—the neural network construction module uses an SGDM optimizer to introduce a momentum term, effectively mitigating gradient oscillations during small-sample training and accelerating convergence. Simultaneously, an L2 regularization term is introduced into the loss function to suppress overfitting by limiting excessive weights. The synergistic effect of these two measures ensures that the system maintains good generalization performance even with limited training samples, guaranteeing the reliability of the decision results. The model training module employs an equal-interval method to construct training samples. This involves determining the value interval based on the optimal and worst values ​​of each indicator, generating multiple virtual samples, and assigning corresponding decision values. This effectively expands the training sample set, covering the complete gradient from optimal to worst, providing ample data support for deep learning model training and solving the practical problem of obtaining large amounts of historical samples in emergency repair. The decision module outputs decision values ​​for each solution. Decision-makers can intuitively compare the merits of each solution through numerical values, avoiding the ambiguity of qualitative evaluation. This provides clear and explicit decision-making basis for emergency repair commanders, facilitating the rapid selection of the optimal solution for implementation.

[0026] Beneficial effects: This invention achieves scientific optimization of power emergency repair engineering schemes through dynamic cost accounting throughout the entire process, construction of a multi-dimensional indicator system, and intelligent decision-making based on an improved convolutional neural network, thereby improving the accuracy and efficiency of decision-making and yielding significant social and economic benefits. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention.

[0028] Figure 2This is a cost breakdown diagram for power emergency repair projects according to the present invention.

[0029] Figure 3 This is a diagram of the direct engineering cost calculation model of the present invention.

[0030] Figure 4 This is a diagram illustrating the cost calculation model for power emergency engineering measures according to the present invention.

[0031] Figure 5 This is the input-output relationship diagram of the present invention.

[0032] Figure 6 This is the CNN structure diagram of the present invention.

[0033] Figure 7 This is a diagram showing the convergence of the model in this invention.

[0034] Figure 8 This is a graph showing the decrease in the loss function of the training and test sets in this invention.

[0035] Figure 9 This is a radar chart showing the performance of the three schemes of the present invention on various indicators.

[0036] Figure 10 This is a graph showing the fitting results of different models of this invention on the training sample set. Detailed Implementation

[0037] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0038] Example 1: This embodiment provides a decision-making method for power emergency repair engineering schemes; such as... Figure 1 As shown, it includes the following steps: S1: Construct a full-process cost accounting model for power emergency repair projects. The full-process cost accounting model includes a direct engineering cost accounting sub-model and an emergency measures cost accounting sub-model, which are used to calculate the cost of alternative repair schemes to obtain direct engineering costs and emergency measures costs. Among them, at least one dynamic correction factor is introduced in the cost accounting to adapt to the uncertainty of the repair scenario. S2: Establish a decision-making indicator system for power emergency repair engineering schemes. The indicator system includes three criterion layers: economic benefits, repair progress, and on-site safety. Each criterion layer has multiple operational level indicators, and the positive or negative attributes of each operational level indicator are clearly defined. S3: Obtain multiple alternative solutions to be decided, calculate the cost data of each alternative solution according to the cost accounting model, and determine the value of each alternative solution on each operational level indicator according to the indicator system to form a solution indicator vector. S4: Construct a scheme decision model based on an improved convolutional neural network, use the SGDM optimizer to optimize the network parameters, and introduce an L2 regularization term into the loss function; S5: Construct a training sample set. The training samples are obtained by equally dividing the value range of each indicator. Each training sample corresponds to a decision value. The improved convolutional neural network is trained using the training sample set. S6: Input the index vectors of the alternative solutions into the trained improved convolutional neural network to obtain the decision values ​​of each solution, and select the optimal emergency repair solution based on the magnitude of the decision values.

[0039] The following is a detailed explanation of each step: S1: Construction of a Full-Process Cost Accounting Model for Power Emergency Repair Projects S1.1 Cost Composition of the Entire Power Emergency Repair Project Power emergency repair projects differ from routine power grid maintenance work, characterized by urgency, coordination, specialization, and high risk, primarily in the following aspects: (1) When a power equipment or line fails, a rapid response and quick repair are required; (2) Requires professional capabilities in rapid diagnosis, emergency repair and power restoration of power systems; (3) It requires working in complex environments such as high-voltage electrical equipment and severe weather, which poses a high safety risk; (4) It requires the coordinated efforts of multiple departments, units and external resources to quickly allocate human, material and financial resources.

[0040] Therefore, the cost of power emergency repair projects includes not only the direct engineering costs included in routine maintenance projects, but also the cost of emergency repair measures, specifically as follows: Figure 2 As shown.

[0041] S1.2 Construction of Sub-item Cost Accounting Model S1.2.1 Direct Engineering Cost Calculation Model (1) Construction machinery usage fee The main factors affecting machinery usage fees include: the type and model of machinery and equipment determined by the initial emergency repair plan; the machinery's shift work quota; the owner's schedule and quality requirements; the work shift system and construction time determined by the nature of the emergency repair task; and the unit price of construction machinery, based on the unit price published by the power quota management agency or the information price approved by the owner when the project actually occurs (e.g., the unit price of cranes).

[0042] (2) Material costs Material costs include two parts: unpriced material costs and priced material costs. (3) Labor costs Labor costs are mainly determined by the number of personnel and the unit price of labor. The personnel allocation type and quantity can be calculated based on the amount of machinery and equipment used and the allocation of emergency repair work areas; the unit price of labor can be determined based on the unit price of labor published by the power quota management agency or the actual unit price of labor actually incurred in the project and approved by the owner.

[0043] The specific principles and process of the direct engineering cost calculation model are as follows: Figure 3 As shown.

[0044] S1.2.2 Emergency Engineering Measures Cost Calculation Model The calculation process for emergency engineering measures costs consists of three parts: "project overview information, cost identification and definition, and cost calculation and generation," which correspond to the three stages of project cost estimation: "preliminary cost estimation of measures, cost statistics of measures, and budget submission of engineering measures." The specific principles and procedures are as follows: Figure 4 As shown. The "preliminary cost estimation of measures" stage is often used for "pre-decision" of emergency repair project construction plans; the "cost statistics of measures" and "budget report of engineering measures" stages are often used for "post-evaluation" of emergency repair projects and as the basis for "pre-decision" of the next stage of repair, that is, the engineering cost data is promptly archived in the database for reference in the next stage of decision-making.

[0045] From a cost aggregation perspective, emergency engineering measures costs consist of emergency dispatch and coordination fees, safety and civilized construction and protection measures fees, and additional construction work fees. From a cost accounting perspective, emergency engineering measures costs are similar to direct engineering costs, consisting of machinery costs, material costs, and labor costs. The accounting models for these various costs are consistent with those for direct engineering costs, and will not be elaborated further here.

[0046] S2: Establish a decision-making indicator system for power emergency repair engineering plans. The construction of a decision-making indicator system for power emergency repair engineering schemes should take the multidimensional needs of power system fault repair and emergency management as its core objectives, and adhere to the principles of systematicness, scientificity, and operability. It should start from three dimensions: economic benefits, repair progress, and on-site safety, deeply integrating the characteristics of the power industry and emergency management practices. Specifically, the economic benefits dimension focuses on cost control and value balance, reflecting the dynamic relationship between resource input and power restoration benefits by quantifying repair costs, returns, and restoration speed. The repair progress dimension revolves around timeliness and technological adaptability, strengthening the supporting role of rapid response and technological iteration in repair efficiency through fault repair time analysis and technical equipment update efficiency assessment. The on-site safety dimension is based on risk prevention and emergency capability building, with safety accident prevention, personnel training coverage, and the effectiveness of plan implementation as its core, constructing a closed-loop management system of pre-event prevention, in-event control, and post-event optimization.

[0047] To ensure that the constructed decision-making indicator system is both comprehensive and concise, and to avoid information overlap and redundancy among indicators, this study uses a systematic three-stage process to complete the initial selection, screening, and final determination of indicators: Preliminary selection phase (literature analysis and industry standard review): By extensively reviewing domestic and international literature on power emergency management, engineering project decision-making and power grid reliability evaluation, and by conducting in-depth research on normative documents such as the "Standards for Budget Preparation and Calculation of Power Grid Maintenance Projects" and the "Safety Production Management Measures" issued by the State Grid Corporation of China and the industry, a pool of candidate indicators covering multiple dimensions such as economy, technology, safety and timeliness has been initially compiled.

[0048] Screening Phase (Delphi Method and Correlation Analysis): The Delphi Method was used to organize expert consultations. Fifteen experts from power grid company operations, emergency repair frontline, engineering cost management, and university research institutes were invited to score the importance and operability of the initially selected indicators and to collect supplementary opinions. Based on the consultation results, indicators with low consensus and poor measurability were eliminated.

[0049] Determining the Stage (Structuralization and Attribute Definition): Combining expert feedback and theoretical analysis, the selected indicators are structured according to the "target layer - criterion layer - operation layer" to form a hierarchical system with "economic benefits", "repair progress" and "on-site safety" as the core criteria. Based on the contribution of the indicators to the merits of the solution, their "positive" or "negative" attributes are clarified.

[0050] Through the above process, the final decision-making system, which includes 14 indicators (as shown in Table 1), not only fully reflects the core requirements of the coordinated optimization of "cost-timeliness-safety" in power emergency repair projects, but also ensures the relative independence and representativeness of each indicator through a strict screening mechanism.

[0051] Table 1. Decision-making index system for emergency repair plans

[0052] The constructed power emergency repair plan decision-making indicator system has a multi-dimensional and multi-level structure. Each level of indicator not only covers the core objective of power emergency repair projects—"short-term and efficient power restoration"—but also considers long-term operation and maintenance capability improvement. Through the combination of quantitative and qualitative indicators, a decision support framework covering the entire repair process and multiple elements is formed. The indicator system takes economic benefits, repair progress, and on-site safety as its three core principles. Through economic benefit indicators such as "repair cost," "repair revenue," and "power restoration speed," it dynamically balances resource input and power restoration benefits to ensure the economic rationality of the repair plan. Progress indicators such as "fault repair time chain" and "equipment technology update rate" quantify the timeliness and technical adaptability of repairs, reflecting the direct impact of repair efficiency on power supply reliability. Through safety indicators such as "accident incidence rate," "safety training coverage rate," and "emergency plan implementation effectiveness," it constructs a risk prevention and emergency response capability assessment framework, strengthening the closed-loop safety management of the entire repair process. The indicators at each level are further refined at the operational level, including quantitative parameters such as direct engineering costs and emergency repair time difference ratios, as well as qualitative dimensions such as the evaluation of the effectiveness of the contingency plan and the improvement of technical equipment efficiency. This forms a decision-making logic of "cost-timeliness-safety" synergistic optimization, which can provide a decision-making basis that takes into account both scientificity and operability for the selection of solutions in complex environments.

[0053] S3: Obtain multiple alternative solutions to be decided, calculate cost data, and construct a solution index vector. This step calculates the cost data for each alternative scheme based on the cost accounting model in step S1, and determines the values ​​of each alternative scheme on each operational level indicator according to the indicator system in step S2, thus forming a scheme indicator vector. Specific case data will be given in the examples below.

[0054] S4: Construct a solution decision model based on an improved convolutional neural network The constructed power emergency repair engineering scheme decision model is used to determine the nonlinear relationship between the decision value Y of the alternative scheme and the decision index system X (14 influencing factor indicators): (10) Because the units of the 14 indicators are difficult to standardize, the decision value Y cannot be directly calculated from all the indicators. Therefore, a deep learning model is used to fit the decision value. The input-output relationship is as follows: Figure 5 As shown.

[0055] Figure 6 A schematic diagram of CNN used to make decisions on power emergency repair engineering schemes is shown. The decision value Y is fitted with 14 indicators in Equation (10) and passed through the input layer (with 14 neurons), the convolutional layer (with 14 convolutional kernels, the kernel size is 3×1), the activation layer (using the ReLU function), the fully connected layer 1 (196 connections), the fully connected layer 2 (196 connections) and the output layer in sequence.

[0056] The formulas for each layer of a convolutional neural network are as follows: (1) Input layer, convolutional layer and activation layer First, 14 metrics are input into the convolutional layer of the convolutional neural network, and then an activation layer is connected after the convolutional layer. Equation (11) means that the functions of the convolutional layer and the activation layer are combined, that is, the convolution operation is performed first, and then the activation is performed. It is a 3×1 convolution kernel. For bias, This represents the output after applying the ReLU function to the convolutional layer. Equation (12) describes the ReLU function.

[0057] (11) (12) (2) Two fully connected layers and output layer (13) (14) The value output by the activation layer is input into the two fully connected layers. In equation (13), These are the weights of the first fully connected layer; The bias is set for the first fully connected layer; For activation functions; This is the output of the first fully connected layer. In equation (14), These are the weights of the second fully connected layer; This is the bias for the second fully connected layer.

[0058] The improved convolutional neural network uses the SGDM optimizer, which selects a mini-batch of samples instead of the entire sample and uses gradient descent to update the weight and bias parameters of the convolutional neural network, aiming to solve the problem of poor fitting performance with random mini-batch samples. A momentum factor is introduced, subtracting the weighted sum of the current and previous gradients when updating the weight and bias parameters, thereby reducing their impact on the weight and bias parameters. An L2 regularization term is also added, with the relevant formulas as follows: (15) (16) (17) (18) (19) (20) (twenty one) Equation (15) represents the loss function of the improved convolutional neural network after adding an L2 regularization term. First, s samples are selected from the training set, including... ,in ( The corresponding target is the cost of power transmission and transformation projects. , and These represent the weights and biases for improving convolutional neural networks; For the total loss function, The loss function without regularization. For L2 regularization terms, is the weight decay coefficient or regularization coefficient. Equations (16) to (18) describe the process of updating weights using the SGDM optimizer with L2 regularization. In equation (16) It is the gradient of the weight parameters. For the number of iterations, It is the gradient of the total loss function with respect to the weight parameters (including L2 regularization); This represents the gradient of the loss function without regularization with respect to the weight parameters; This represents the gradient of the regularization term with respect to the weight parameters. In equation (17), For the number of iterations, For the weight parameter of the first The cumulative momentum of the next iteration; For momentum parameters, For the weight parameter of the first The cumulative momentum of the next iteration Let be the learning rate. In equation (18), For the first Weight parameters for the next iteration For the first The weight parameters for the next iteration. Equations (19) to (21) represent the process of updating the bias using L2 regularization of the SGDM optimizer. Since L2 regularization is not required for the bias, no L2 regularization term is set in equation (19), and equation (10)... It is the gradient of the bias parameters. For the number of iterations, It is the gradient of the total loss function with respect to the bias parameters (excluding L2 regularization). Let represent the gradient of the loss function without regularization terms with respect to the bias parameters. In equation (20), For the first bias parameter The cumulative momentum of the next iteration The number of iterations; For momentum parameters, For the first bias parameter The cumulative momentum of each iteration. In equation (21), For the first The bias parameters for the next iteration. For the first The bias parameters for the next iteration.

[0059] By treating the 14 decision indicators as a one-dimensional "sequence," each convolutional kernel can actively learn the local synergistic or antagonistic relationships between adjacent indicators within a small range (e.g., indicators C6, C7, and C8, all belonging to the "repair progress" dimension) during the sliding process. This aligns better with the decision-maker's thinking when comprehensively evaluating multiple indicators. Furthermore, the weight-sharing mechanism significantly reduces model parameters, and combined with L2 regularization, it effectively suppresses overfitting and improves the model's generalization performance with limited training samples. Therefore, although CNNs are traditionally used for image recognition, their advantage in mining local features in structured data makes them highly suitable for solving the multi-dimensional, heterogeneous indicator collaborative decision-making problem faced in this study.

[0060] S5: Construct a training set and train the improved convolutional neural network using the training sample set. S5.1: Constructing the training set First, construct a training set for the decision-making indicator system. X Assume that for a certain power emergency repair project, there exists... One alternative The value of the i-th alternative on the 14 decision indicators is denoted as . For indicators j This allows us to obtain the optimal value of the indicator among all alternative solutions. and worst value : (8) (9) It can be seen that the range of values ​​for a positive indicator is... For negative indicators, the range of values ​​is: .

[0061] Divide the value range of the indicator into equal parts. You can get a portion. Interval points. For example, for the first division point, when the index j As a positive indicator, its value is... When the indicator j It is a negative indicator, and its value is... Based on this, we can construct A sample is used as the training set for the decision-making indicator system.

[0062] Secondly, construct a training set of decision results. YTo ensure the intuitiveness and comparability of the decision results, this embodiment sets the range of values ​​for the decision results as follows: The constructed training set for decision results should be consistent with the training set for decision indicators; therefore, the value range of the decision results should be divided into equal parts. You can get a portion. Interval points For example, the optimal set of indicators in the training set of decision indicators. The corresponding decision outcome interval is 1, and the worst-case indicator set in the decision indicator training set. The corresponding decision result interval is 0.

[0063] S5.2: Train the improved convolutional neural network using the aforementioned training sample set. Network training parameters are set as follows: maximum number of iterations 100, number of samples traversed 5. Bayesian optimization is used to optimize the hyperparameters, with the learning rate ranging from [0.01, 1], the stochastic gradient descent momentum ranging from [0.8, 0.98], and the L2 regularization strength ranging from [1 × 10⁻⁶]. - ¹ 0 The loss function is the loss function with L2 regularization added (Equation (15)). The SGDM optimizer is used to update the weights and biases. A momentum term is introduced during the update process to reduce gradient oscillations.

[0064] The network is trained using the m+1 training samples constructed in step S5.1, and the average absolute percentage error is used as the fitness value.

[0065] S6: Input the index vectors of the alternative solutions into the trained improved convolutional neural network to obtain the decision values ​​of each solution, and select the optimal emergency repair solution based on the magnitude of the decision values. The index vectors of the alternative solutions are input into the trained improved convolutional neural network to obtain the decision values ​​of each solution, and the optimal emergency repair solution is selected based on the magnitude of the decision values.

[0066] The following uses an ice storm relief project as an example to further illustrate the embodiments of the present invention.

[0067] Taking a power emergency repair project during an ice storm as the application background, this study focuses on the tower erection and line stringing operation following a tower collapse at a project site. The site is located at a high altitude, experiencing persistently low temperatures in the mountains, and is subject to freezing rain, exhibiting typical characteristics of a complex emergency repair scenario. Three alternative repair plans are proposed for this project: Option 1: A suspended gantry tower is constructed, primarily using external guy wires and secondarily internal guy wires. By analyzing the structure and function of the external guy wires on the internal suspended gantry, the external guy wires are used for hoisting and installation. Precise calculations of the tension and direction of the external guy wires are required. This option is suitable for relatively open construction environments, but requires a longer time to disassemble and reassemble the tower, placing high demands on hoisting equipment and personnel.

[0068] Option 2: A suspended gantry tower is constructed, primarily using internal guy wires and secondarily using external guy wires. By analyzing the structure and function of the internal guy wires in the internal suspended gantry, these wires are used to fix each component in a suitable position. Precise control of the tension and direction of the internal guy wires is required. This option is suitable for construction in confined spaces, but it is more challenging and requires highly skilled engineers and operators.

[0069] Option 3: Deconstruct and reassemble the tower using a ground-based gantry crane. This involves hoisting the demounted tower to the ground, rebuilding a support structure on the ground, and then erecting the tower again. This requires large hoisting equipment. The hoisting process is relatively safe, but the construction period is longer, making it suitable for more complex construction environments.

[0070] This embodiment is implemented according to the following steps: Step 1: Construct a full-process cost accounting model and perform cost accounting for alternative solutions. (1) Direct Engineering Cost Calculation Sub-model Direct project costs include construction machinery usage fees, material costs, and labor costs.

[0071] Construction machinery usage fee: Determined based on the amount of machinery used and the unit price per machine shift. The amount of machinery used is calculated based on the project volume, machinery shift operating rate, work shift system, and planned construction period. For example, if the excavation volume is 800 m³, the machinery shift operating rate is 40 m³ / shift, there are 2 shifts per day, and the planned construction period is 2 days, then the machinery usage is 5 units / day. The unit price per machine shift is determined based on the information price published by the power quota management agency, such as a crane unit price of 5200 yuan / shift.

[0072] Material costs: This includes priced material costs and unpriced material costs. Priced material costs are based on the electricity quota, such as 18,000 yuan per day. Unpriced material costs include fuel, special cables, steel, etc., and their costs are calculated based on material usage and the budgeted unit price. Fuel costs are adjusted using dynamic fuel consumption correction factors, including speed coefficient, machine condition coefficient, and condition coefficient. For example, if the basic fuel consumption quota for machinery is a certain value, the actual fuel consumption is obtained after adjustments based on driving time, operating time, machine condition coefficient, and condition coefficient, and then multiplied by the fuel price to arrive at the fuel cost.

[0073] Labor costs: Determined based on the number of personnel and the unit price of labor. Personnel allocation is determined based on the amount of machinery used and the needs of the work area. For example, Option 1 allocates 6 technicians and 12 general workers, with a unit price of 600 yuan / day for technicians and 400 yuan / day for general workers.

[0074] (2) Emergency Measures Cost Calculation Sub-model Emergency response costs include emergency dispatch and coordination costs, safety and civilized construction and protection costs, and additional construction work costs. Each cost is calculated based on the consumption and unit price of labor, materials, and machinery. For example, emergency dispatch and coordination costs are 15,000 yuan per day, safety and civilized construction and protection costs are 20,000 yuan per day, and additional construction work costs are 7,500 yuan per day.

[0075] During the calculation process, a dynamic correction factor is introduced to adapt to the uncertainties of emergency repair scenarios. In this embodiment, the terrain complexity adjustment coefficient is set to 1.15 (mountainous area), and the machine condition coefficient and condition coefficient are determined based on the actual machine conditions and altitude conditions.

[0076] Based on the above model, cost accounting was performed on the three alternative schemes to obtain the direct engineering costs and contingency measures costs for each scheme.

[0077] The key parameters required for cost accounting of the three alternative options are shown in Table 2.

[0078] Table 2 Key parameters for cost accounting of three alternative options

[0079] Based on the parameters listed in Table 2, the full-process cost calculation results of the three alternative schemes were obtained by using the established accounting model, as shown in Table 3.

[0080] Table 3 Cost Calculation Results of Three Alternative Schemes

[0081] Step 2: Establish a decision-making indicator system and determine the indicator values ​​for each option. A decision-making indicator system was established, comprising three criterion layers: economic benefits, repair progress, and on-site safety. Each criterion layer has several operational level indicators, totaling 14 items, and the positive or negative attributes of each indicator are clearly defined. The specific indicators and the actual values ​​of the three schemes are shown in Table 4.

[0082] Table 4 Decision-making index values ​​for power emergency repair construction plans

[0083] Based on the cost data calculated in step 1, and in conjunction with the above indicators, the values ​​of each scheme on the 14 indicators are determined, forming a scheme indicator vector.

[0084] Step 3: Construct the training sample set Using the maximum and minimum values ​​of all candidate schemes in the power emergency repair project decision index values ​​as boundaries, and based on the distribution of each index of the three repair schemes, the value range of each index is divided into 10 equal parts, that is, the schemes are divided into 11 levels from best to worst. Ranked from best to worst, 11 training samples are obtained, with each level corresponding to a target value. This forms the target vector Y=[1.0,0.9,0.8,0.7,0.6,0.5,0.4,0.3,0.2,0.1,0], forming the decision index values ​​of all training samples, as shown in Table 5.

[0085] Table 5 Decision Index Values ​​of Power Emergency Repair Construction Scheme Training Sample

[0086] Step 4: Build and train an improved convolutional neural network An improved convolutional neural network is constructed, consisting of an input layer, convolutional layers, activation layers, fully connected layers, and an output layer. The input layer has 14 nodes, corresponding to 14 metrics; the output layer has 1 node and outputs the decision value. The convolutional layers use 14 convolutional kernels of size 3×1, and the ReLU activation function is used; there are two fully connected layers, each with 196 connections.

[0087] Network training parameters are set as follows: maximum number of iterations 100, number of samples traversed 5. Bayesian optimization is used to optimize the hyperparameters, with the learning rate ranging from [0.01, 1], the stochastic gradient descent momentum ranging from [0.8, 0.98], and the L2 regularization strength ranging from [1 × 10⁻⁶]. - ¹ 0 [0.01]. The loss function is the loss function with L2 regularization added, which is the average loss of all training samples plus the L2 regularization penalty multiplied by the regularization coefficient. The SGDM optimizer is used to update the weights and biases, and a momentum term is introduced during the update process to reduce gradient oscillations.

[0088] The network was trained using the 11 training samples constructed in step 3, with the mean absolute percentage error used as the fitness value. The convergence during training is as follows: Figure 7 As shown, the loss function decrease curve is as follows: Figure 8 As shown in the figure, the model decreases smoothly and converges synchronously on both the training and test sets, indicating that overfitting has not occurred.

[0089] Step 5: Solution Decision The index vectors of the three alternative schemes are input into the trained improved convolutional neural network to obtain the decision values ​​of each scheme, as shown in Table 6: Table 6 Decision Results

[0090] The performance of the three schemes on various indicators is as follows: Figure 9 As shown, Option 1 has the shortest power restoration time (4 days) and the shortest fault response and emergency repair start-up time (both 12 hours). Furthermore, the efficiency improvement rate of the new equipment technology (93.43%) and the effectiveness of the emergency plan execution (89 points) are both at a high level. Although its direct engineering costs and emergency measures costs are slightly higher (71,300 yuan / day and 42,500 yuan / day respectively), the rapid power restoration significantly reduces the economic losses caused by power outages. It also performs excellently in economic benefit indicators such as the fixed asset value transformation ratio (91.78%) and return on investment (98.13%), demonstrating a reasonable balance between cost and timeliness. In comparison, Option 3 has the lowest decision value (0.4009, Level VII). Although its direct engineering costs and emergency measures costs are the lowest (58,535 yuan / day and 39,500 yuan / day, respectively), the power restoration time is as long as 10 days, and indicators such as the fixed asset value change ratio (86.10%) and return on investment (84.55%) are significantly lagging behind, indicating that simply reducing costs cannot compensate for the overall loss of benefits caused by insufficient repair efficiency. Option 2 (0.6215, Level IV) is in the middle between cost and efficiency. Its emergency measures costs are relatively high (48,000 yuan / day), but the application rate of new technology equipment (30.55%) and the improvement rate of repair efficiency (95.17%) do not show a significant advantage, resulting in a middle overall score. Overall, the constructed power emergency repair engineering scheme decision model, by balancing economy, timeliness, and safety, verifies the applicability of Option 1 in complex emergency scenarios, and reveals the importance of avoiding overemphasis on a single dimension and neglecting the synergy of multiple indicators in repair scheme decision-making.

[0091] according to Figure 9 The radar chart and the decision results in Table 6 clearly show that Scheme 1 is significantly better than Schemes 2 and 3 in terms of key timeliness indicators—power restoration time (C5) and emergency response time (C6). This difference mainly stems from the technical characteristics and applicable scenarios of the three emergency repair schemes themselves, and directly determines the superiority or inferiority of their overall decision values.

[0092] The reason why Option 1 (suspended pole with external guy wire) performs best in C5 and C6 is that it uses external guy wires for hoisting, which allows for a large hoisting range and high flexibility in a relatively open construction environment, significantly improving equipment assembly and positioning efficiency. This enables personnel and machinery to quickly begin work upon arrival on site (excellent C6 index), and the entire tower erection process is smoother, thus greatly shortening the total time from the start of emergency repairs to power restoration (excellent C5 index). Its technical approach essentially aligns with the core requirement of "speed" in emergency repairs.

[0093] The reason why Option 2 (suspended pole with internal guy wire) performs worse in C5 and C6 is that it is more suitable for operations in confined spaces. Although the internal guy wire system enhances stability, its installation, tension adjustment, and synchronization control are more complex, requiring higher technical skills. This, to some extent, prolongs the preparation time before construction and the coordination time during the process, resulting in a slightly lower response and repair efficiency (C5, C6) than Option 1.

[0094] Option 3 (ground-mounted tower crane) performs the worst in C5 and C6 because it requires the tower crane components to be completely grounded and disassembled and reassembled on the ground. This process involves extensive ground assembly and multiple lifting operations by large cranes, making it the most cumbersome and time-consuming. Therefore, although it is relatively safe when operating on the ground, it incurs huge time costs, resulting in a power restoration time (C5) of up to 10 days and the slowest emergency response (C6).

[0095] The decision-making system constructed in this model highly values ​​the "repair progress" dimension, as rapidly restoring power supply is the primary goal of emergency repairs. Indicators C5 (power restoration time) and C6 (emergency response time), as core representatives of this dimension, are given key weights by the model. The absolute advantage of Option 1 in these two core timeliness indicators is the fundamental reason for its highest decision value (0.8071). This indicates that in real emergency repair decisions, investing higher direct costs (such as Option 1) to shorten the power outage time by several days is usually worthwhile, because the socio-economic losses caused by the power outage far outweigh the increased engineering costs of the repair itself. Conversely, although Option 3 has the lowest cost, its excessively long repair cycle leads to it being judged as the worst option in the comprehensive evaluation. This strongly verifies that the model in this embodiment does not simply perform cost accounting, but rather conducts a scientific "cost-timeliness" trade-off, and its decision results are highly consistent with the actual priorities of power grid emergency management.

[0096] Step 6: Disturbance Analysis To further verify the model's sensitivity to key indicators, based on the data in Table 6, each indicator was varied by ±20%, and the changes in the decision values ​​of the three schemes were calculated. The results are shown in Table 7 (only significant disturbances with a change of more than 5% are shown).

[0097] Table 7 Results of Disturbance Analysis of Key Indicators

[0098] As shown in Table 7, changes in power restoration time (C5) have the most significant impact on the decision value of the proposed solutions. When the power restoration time of Solution 1 increases by 20%, its decision value decreases by 11.75%; while the same change in the same indicator for Solution 3 leads to a decrease in the decision value of 12.40%, confirming the core role of timeliness in emergency decision-making. Meanwhile, the core advantage of Solution 1 relies on multi-dimensional synergy: ±20% disturbances in the fixed asset value ratio (C3) and the effectiveness of the safety plan (C14) cause fluctuations in the decision value of approximately ±5.6%, indicating that its high score (0.8071) stems not only from rapid response capabilities but also from the comprehensive performance of asset preservation and plan execution. In contrast, although Solution 3 has the lowest cost, a prolonged power restoration time or a decline in the effectiveness of the safety plan will further worsen its score (fluctuation of 7.93%~12.40%), highlighting that simply reducing cost cannot compensate for shortcomings in efficiency and safety. It is worth noting that the decision value of Scheme 2 is highly sensitive (±9.8%) to the efficiency of new technologies (C10) and the cost of emergency measures (C1), suggesting that the gap with Scheme 1 can be narrowed by optimizing the cost-effectiveness of technology application. Overall, the disturbance analysis verifies the model's comprehensive consideration of the timeliness, safety, and economy of power emergency repair projects, and the significant disturbance directions of each indicator are consistent with the actual needs of the project, indicating that the model effectively captures the complex nonlinear relationships between multiple objectives.

[0099] Step 7: Model Performance Validation and Analysis To verify the superiority of the proposed improved CNN model (denoted as Improved-CNN) in solving small-sample, multi-index decision-making problems, this study designed a comparative experiment. The performance of Improved-CNN was compared with that of several classic and cutting-edge prediction and classification models on the same training and test sets (see Tables 4 and 5). The selected comparison models included: Multiple linear regression (MLR): As a baseline linear model, it is used to compare the gains of nonlinear models.

[0100] Support Vector Machine (SVM): A classic machine learning algorithm that performs robustly with small sample sizes.

[0101] Backpropagation Neural Network (BPNN): A classic fully connected neural network used to compare the advantages of CNN structures in capturing local features and preventing overfitting.

[0102] Conventional Convolutional Neural Network (CNN): A basic CNN version without SGDM optimizer and L2 regularization, used to evaluate the contribution of the two improvements (SGDM+L2) in this embodiment.

[0103] All comparison models used the same inputs (14 metrics) and outputs (decision values). The fitting accuracy of each model on the training set was calculated, and the results are as follows: Figure 10 As shown.

[0104] according to Figure 10 It can be seen that the improved CNN model proposed in this embodiment has the closest fitting result to the target value, especially in extreme levels. The error is only 0.9% in the best level (Level I) and 2.1% in the worst level (Level XI), which is significantly better than other models (such as SVM with an error of 2.5% in Level I and 4.5% in Level XI; and BP with an error of 3.8% in Level I). Although there are slight fluctuations in intermediate levels (such as Level VII with a target of 0.4) (error of 4.75%), its overall curve is smooth and the degree of deviation is much lower than that of linear regression (10% error in Level VII) and conventional CNN (abnormal fluctuations in Levels IV-V). Figure 7 The convergence curve analysis shows that the model's error gradually decreased during training without drastic oscillations, indicating that the parameter optimization process was stable and the hyperparameters (such as learning rate and regularization strength) were set reasonably. Overall, the improved CNN, by introducing the SGDM optimizer, effectively alleviates the gradient oscillation problem in small-sample training. Combined with L2 regularization to suppress overfitting, it can accurately capture the complex nonlinear relationships among 14 heterogeneous indicators, verifying the model's effectiveness in solving multi-dimensional collaborative decision-making problems.

[0105] To quantitatively evaluate the performance of each model, the mean absolute error (MAE) and root mean square error (RMSE) of each model on the training set were calculated, and the results are shown in Table 8. It can be seen that the Improved-CNN proposed in this embodiment significantly outperforms the comparative models in all error metrics. Its MAE (0.021) and RMSE (0.028) are reduced by approximately 41.7% and 39.1% compared to the second-best SVM model, respectively. Furthermore, in the optimal and worst-case levels representing extreme cases, the prediction error of Improved-CNN is also much smaller than that of other models. These quantitative results fully demonstrate that by introducing the SGDM optimizer and L2 regularization strategy, the improved CNN effectively enhances the fitting accuracy and generalization ability in scenarios with small samples and multiple metrics, and its performance advantage is statistically significant.

[0106] Table 8 Comparison of fitting errors of different models on the training set

[0107] Example 2: This embodiment provides a power emergency repair engineering scheme decision system, which includes: The cost accounting module is used to construct a full-process cost accounting model for power emergency repair projects, including direct engineering costs and emergency measures cost accounting sub-models. It performs cost accounting for alternative repair schemes and introduces at least one dynamic correction factor to adapt to the uncertainty of the repair scenario. The indicator system construction module is used to establish a decision-making indicator system for power emergency repair engineering schemes, which includes three criterion layers: economic benefits, repair progress, and on-site safety. Each criterion layer has several operational level indicators, and the positive or negative attributes of each operational level indicator are clearly defined. The scheme indicator acquisition module is used to acquire multiple alternative emergency repair schemes to be decided, call the cost accounting module to obtain the cost data of each scheme, and then determine the values ​​of each scheme's indicators at each operational level based on the decision indicator system to form a scheme indicator vector. The neural network building module is used to construct an improved convolutional neural network scheme decision model containing an input layer, convolutional layer, activation layer, fully connected layer and output layer. It uses the SGDM optimizer to optimize network parameters and introduces an L2 regularization term into the loss function. The model training module is used to construct a training sample set obtained by equally dividing the index value range, each training sample corresponding to a decision value, and to train the improved convolutional neural network using the training sample set. The scheme decision module is used to input the index vectors of each alternative scheme into the trained scheme decision model, obtain the decision value of each scheme, and select the optimal emergency repair scheme based on the magnitude of the decision value.

[0108] This system provides a scientific, efficient, and reliable solution for power emergency repair engineering decision-making through technological innovations such as full-process dynamic accounting, multi-dimensional collaborative evaluation, intelligent neural network modeling, and small-sample generalization enhancement, which is conducive to improving the power grid's disaster resistance capabilities and operation and maintenance efficiency.

[0109] In specific implementation, the working principle, control process and technical effects of the power emergency repair engineering scheme decision system provided in this embodiment of the invention are the same as those of the power emergency repair engineering scheme decision method in the foregoing embodiment, and will not be repeated here.

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A decision-making method for power emergency repair engineering schemes, characterized in that, Includes the following steps: 1) Construct a full-process cost accounting model for power emergency repair projects. The full-process cost accounting model includes a direct engineering cost accounting sub-model and an emergency measures cost accounting sub-model, which are used to calculate the cost of alternative repair schemes to obtain direct engineering costs and emergency measures costs. Among them, at least one dynamic correction factor is introduced in the cost accounting to adapt to the uncertainty of the repair scenario. 2) Establish a decision-making indicator system for power emergency repair engineering schemes. The indicator system includes three criterion layers: economic benefits, repair progress, and on-site safety. Each criterion layer has multiple operational level indicators, and the positive or negative attributes of each operational level indicator are clearly defined. 3) Obtain multiple alternative solutions to be decided, calculate the cost data of each alternative solution according to the cost accounting model, and determine the value of each alternative solution on each operational level indicator according to the indicator system to form a solution indicator vector. 4) Construct a scheme decision model based on an improved convolutional neural network, use the SGDM optimizer to optimize the network parameters, and introduce an L2 regularization term into the loss function; 5) Construct a training sample set. The training samples are constructed by synchronously dividing the value range of each indicator into equal parts. That is, for each indicator, the value range is determined according to its best and worst values ​​and the range is divided into several levels. The interval points of the same level of each indicator are taken to form a training sample. Each training sample corresponds to a decision value. The improved convolutional neural network is trained using the training sample set. 6) Input the index vectors of the alternative schemes into the trained improved convolutional neural network to obtain the decision values ​​of each scheme, and select the optimal emergency repair scheme based on the magnitude of the decision values.

2. The decision-making method for power emergency repair engineering schemes according to claim 1, characterized in that, The direct engineering cost accounting sub-model includes construction machinery usage cost accounting, material cost accounting, and labor cost accounting; The construction machinery usage fee is determined based on the amount of machinery used and the unit price of machinery per shift, wherein the amount of machinery used is calculated based on the amount of work, the machinery shift operation rate, the work shift system and the planned construction period. The material cost includes priced material cost and unpriced material cost. The fuel cost in the unpriced material cost is corrected by a dynamic fuel consumption correction factor, which includes a speed coefficient, an engine condition coefficient, and a condition coefficient. The labor cost is determined based on the number of personnel and the unit price of labor.

3. The decision-making method for power emergency repair engineering schemes according to claim 2, characterized in that, The dynamic fuel consumption correction factor is used to correct the basic fuel consumption quota of the machinery in order to calculate the actual fuel consumption of the machinery under emergency repair conditions.

4. The power emergency repair engineering scheme decision-making method according to claim 1, characterized in that, The emergency response cost accounting sub-model is used to calculate emergency dispatch and coordination costs, safety and civilized construction and protection costs, and additional construction work costs. Each cost is calculated based on the consumption and unit price of labor costs, material costs, and machinery usage costs.

5. The decision-making method for power emergency repair engineering schemes according to claim 1, characterized in that, The operational level indicators mentioned in step 2) include: power emergency engineering measures cost, direct engineering cost, change in value of fixed assets before and after emergency repair, return on investment, time required to restore power supply to normal status, time from receiving a fault alarm to the arrival of personnel, materials and equipment on site, time from arrival on site to the start of emergency repair, the difference between actual repair time and planned repair time, the proportion of emergency repairs using the latest technology and equipment, the proportion of improvement in repair efficiency due to the introduction of new equipment or technology, the frequency of unexpected events during emergency repairs, the proportion of emergency repair personnel who have received safety training, the degree of implementation of the emergency plan in actual emergency repairs, and the evaluation score of the effectiveness of the emergency plan after implementation. Among them, the following are negative indicators: power emergency engineering measures cost, direct engineering cost, time required to restore power supply to normal status, time from receiving a fault alarm to the arrival of personnel, materials and equipment on site, time from arrival on site to the start of emergency repair, the ratio of actual repair time to planned repair time, the proportion of repairs using the latest technology and equipment, and the frequency of unexpected events during the repair process; the following are positive indicators: the ratio of fixed asset value change before and after repair, return on investment, the proportion of improvement in repair efficiency due to the introduction of new equipment or technology, the proportion of repair personnel receiving safety training, the degree of implementation of the emergency plan in actual repairs, and the evaluation of the effectiveness of the emergency plan after implementation.

6. The decision-making method for power emergency repair engineering schemes according to claim 1, characterized in that, The method for constructing training samples in step 5) is as follows: For each indicator, the value range of the indicator is determined according to the optimal and worst values ​​in the alternative schemes. The value range is divided into several equal parts to obtain multiple interval points. Each interval point corresponds to an indicator value. Interval points of the same level of each indicator are taken to form a training sample, thereby obtaining multiple training samples with the same number of levels. The corresponding decision values ​​are assigned in order from optimal to worst, and the decision value range is [0,1].

7. The power emergency repair engineering scheme decision-making method according to claim 1, characterized in that, The improved convolutional neural network includes an input layer, a convolutional layer, an activation layer, a fully connected layer, and an output layer. The number of nodes in the input layer of the improved convolutional neural network is equal to the number of indicators in the operation layer, the number of nodes in the output layer is 1, the convolutional layer uses multiple convolutional kernels of fixed size, the activation function is the ReLU function, and there are two fully connected layers.

8. The decision-making method for power emergency repair engineering schemes according to claim 1, characterized in that, The loss function of the improved convolutional neural network is a loss function with added L2 regularization. The loss function with added L2 regularization is the average loss of all training samples plus the L2 regularization penalty term multiplied by the regularization coefficient. When updating the network parameters using the SGDM optimizer, a momentum term is introduced during the update process of weights and biases to reduce gradient oscillations, and overfitting is suppressed by the L2 regularization term.

9. The decision-making method for power emergency repair engineering schemes according to claim 1, characterized in that, The dynamic correction factor includes at least one of the terrain complexity adjustment coefficient, machine condition coefficient, and condition coefficient, and is used to correct the cost of construction machinery or material costs.

10. A decision-making system for power emergency repair engineering schemes, characterized in that, include: The cost accounting module is used to construct a full-process cost accounting model for power emergency repair projects, including direct engineering costs and emergency measures cost accounting sub-models. It performs cost accounting for alternative repair schemes and introduces at least one dynamic correction factor to adapt to the uncertainty of the repair scenario. The indicator system construction module is used to establish a decision-making indicator system for power emergency repair engineering schemes, which includes three criterion layers: economic benefits, repair progress, and on-site safety. Each criterion layer has several operational level indicators, and the positive or negative attributes of each operational level indicator are clearly defined. The scheme indicator acquisition module is used to acquire multiple alternative emergency repair schemes to be decided, call the cost accounting module to obtain the cost data of each scheme, and then determine the values ​​of each scheme's indicators at each operational level based on the decision indicator system to form a scheme indicator vector. The neural network building module is used to construct an improved convolutional neural network scheme decision model containing an input layer, convolutional layer, activation layer, fully connected layer and output layer. It uses the SGDM optimizer to optimize network parameters and introduces an L2 regularization term into the loss function. The model training module is used to construct a training sample set obtained by synchronously dividing the value range of each indicator. For each indicator, the value range is determined according to its best and worst values ​​and the range is divided into several levels. The interval points of the same level of each indicator are taken to form a training sample. Each training sample corresponds to a decision value, and the improved convolutional neural network is trained using the training sample set. The scheme decision module is used to input the index vectors of each alternative scheme into the trained scheme decision model, obtain the decision value of each scheme, and select the optimal emergency repair scheme based on the magnitude of the decision value.

Images